FabricPC

What is FabricPC?

FabricPC is a research library for building and experimenting with predictive coding networks, a brain-inspired alternative to the way neural networks are usually trained. Rather than sending an error signal back across the whole network, each node makes local predictions and learns from its own errors, which makes it a natural fit for recurrent and arbitrary graph architectures.

Models are assembled from three abstractions: nodes (latent state, projection functions, and activations), wires (the edges connecting them), and updates (the inference and local-learning steps). Together these let you build feedforward, recurrent, or skip-connection architectures from modular parts. It runs on JAX for GPU acceleration and node-level automatic differentiation, currently supports linear and convolutional nodes with transformers in progress, and is built to extend with custom node, inference, and initialiser types. It’s MIT-licensed and research-first, so expect APIs to shift ahead of v1.0. To start, use an editable install in a Python 3.12 environment and run the MNIST demo in the examples folder.